Arm unveiled three new pieces of chip IP this week: the C2-Ultra CPU core, the Mali G2-Ultra NX GPU, and the Neoverse CSS N4 datacenter platform. Arm does not manufacture chips. It licenses core designs to phone makers and chipmakers, who then build those designs into their own silicon. That licensing model is why one Arm announcement can end up shaping the processor inside dozens of unrelated phones within a year. Xiaomi has already started that process: its XRING O3 chip, which launched August 24, ships with the G2-Ultra NX GPU built in, according to Chips and Cheese, the independent hardware publication that examined Arm’s disclosures.
For an AI audience, the part of this announcement worth tracking sits inside the GPU. Arm added a matrix accelerator to each shader core in G2-Ultra NX, a block built to run the multiply-accumulate math that neural network inference depends on, and one that can clock up to twice as fast as the GPU’s regular execution units. That is the kind of addition that matters for running models on a phone rather than in a data center. Chips and Cheese flagged a real limit, though: the accelerator handles only INT8 and INT16 arithmetic. It does not support FP8 or BF16, the lower-precision formats that on-device inference stacks have leaned on to balance accuracy against battery life. A GPU built with local inference in mind arrived without the number formats those workloads are moving toward.
Arm’s disclosure record on the CPU side raises separate concerns. The company originally claimed a 15 percent peak and 12 percent average performance gain for C2-Ultra over its predecessor, C1-Ultra. Arm has since told Chips and Cheese, in a correction the publication added to its piece, that the actual maximum IPC gain is 7 percent, and that figure still folds in a larger L2 cache rather than isolating the redesigned core. Chips and Cheese’s own recalculation, which adjusts for the 8.5 percent clock speed bump built into Arm’s benchmark comparison, puts the average uplift closer to 3.2 percent. Separately, Arm says C2-Ultra draws 38 percent less power than C1-Ultra, a figure that blends manufacturing process gains with architectural ones and should not be read as a pure microarchitecture result.
The GPU comparison follows the same pattern. Arm claims up to 14 percent faster games and 24 percent faster ray tracing versus the prior G1-Ultra generation. Chips and Cheese points out the comparison chips ran roughly 11 percent higher clocks, which means most of the non-ray-tracing gain likely comes from frequency rather than new circuitry.
None of these figures come from shipping hardware tested at scale. They are Arm’s own FPGA-simulation results, examined and partly corrected by an outside publication after the fact. Chips and Cheese was direct about why that correction was necessary: Arm’s presentation left out details the outlet had to request by email afterward, a gap that applied to the CPU claims and to the barely-disclosed Neoverse CSS N4 server platform, which Arm covered in a single slide before answering follow-up questions on core counts and cache sizes.
For anyone tracking on-device inference roadmaps, the practical read is narrower than the announcement suggests. G2-Ultra NX brings genuinely new matrix hardware to phones, but only in INT8 and INT16, so a local model pipeline built around FP8 or BF16 quantization will need a separate execution path, or a wait for the next Arm generation, before this silicon becomes useful for it.
Chips and Cheese published this analysis, with a correction from Arm added on September 8, 2026.
